Restructure the cmake cuda process. Remove the policy and always set CMAKE_CUDA_ARCHITECTURE to 30, if not set. This will fail when someone is trying to build with cuda 10.2 or higher, as cuda 10.2 minimum support is 52.

NOTICE: preferred way is to set CMAKE_CUDA_ARCHITECTURE as cmake argument OR in the config file to the real gpu architecture. For example phoenix tesla P100 have the compute capability 60.

Adapt the <only> syntax to branch@organisation/repo according to https://docs.gitlab.com/ee/ci/yaml/#onlyexcept-basic

Remove debugging ls. Activate double accuracy for the numerical tests. Deactivate 256 & 512 grid sizes.